Comparing hybrid and machine learning models for rainfall-induced shallow landslide susceptibility.

Effective risk management for rainfall-induced mass movements necessitates models that simultaneously have predictive power and are interpretable. This study introduces a hybrid model combining machine learning and physical principles to enhance landslide susceptibility predictions under rainfall conditions. By in­tegrating the predictive power of machine learning with an infinite-slope stability approach, we overcome traditional limitations in model parametrization. We utilized static and dynamic key factors such as terrain surface properties, lithological and tectonic setting, land cover/land use and accumulated precipitation to estimate ef­fective cohesion, internal friction angle and soil parameters as latent variables of a Factor-of-Safety model based on the law of Mohr-Coulomb. We compared this approach with established physically-based and empirical models for a landslide/non-landslide point sample of 620 observations in a study area in northern and west­ern Slovenia (400km²) where disastrous rainfall events happened in August and autumn 2023. A spatial block cross-vali­dation yielded median Area under the Receiver Operating Curve values of 0.53 for the physical model, 0.85 for the benchmark empirical model and 0.66 for the hybrid model. Despite inferior model performance in discriminating between landslide and non-landslide points compared to the benchmark, the hybrid model al­lows for the interpretation, mapping and plausibility check of geo­mechanical and hydrological slope stability parameters.